I have been rather occupied of late, yet in my spare moments, words remain my sanctuary. My schedule involves classes starting at 8:30 AM and running until evening self-study ends at 9:00 PM. Occasionally, I even put in “overtime”, not returning home until 10:00 PM. This routine persisted for forty-one consecutive days. Given my mental state under such intensity, returning home to play reaction-based games like osu! was out of the question; I didn’t even have the energy to click through a single chapter of a visual novel.
The only viable activity was to engage in small talk with an LLM (Large Language Model).
I must confess, although I maintain a blog, I have never managed to articulate my views or philosophies (or perhaps simply my “ideas”) in a systematic fashion. I have often thought that this collection of overly broad and disparate viewpoints is difficult to convey through plain narrative. To address this, I attempted various methods, such as using a random event as a cross-section to “slice open” the tangled mess of my thoughts and depict their internal structure.1 Yet, these methods never allowed me to express my inner voice freely. Writing does not flow effortlessly from my pen; I cannot simply write whenever I wish. Ultimately, a blog is meant for an audience, so when the words wouldn’t come, I resorted to writing simple technical posts to garner some search engine traffic.
The advent of LLMs, however, has altered this dynamic. All my thoughts can now receive immediate responses in relative privacy. Why have I not written a blog post in so long? Fatigue is naturally a major factor, but the fact that LLMs satisfy a significant portion of my need for expression and validation—without the need for self-censorship—has diluted my desire to blog during this period.
There is no doubt that composing long-form text is beneficial for the brain. The impact of LLMs on my ability to write at length is likely akin to, if not greater than, the impact micro-blogging (like Twitter or Weibo) had when it first appeared. Micro-blogging habituated people to fragmented expression and immediate feedback, weakening the ability to construct complex arguments and long-form narratives. LLMs go a step further: they not only provide a channel for expression but also directly engage in providing emotional value, acting as a substitute for a “confidant”. When posting on social media, one must consider privacy, controversy, or simply whether the content is worth exposing; the publisher invariably exercises some degree of caution.
LLMs are different. The worst-case scenario for content sent to an API is that it gets used to train the model, not that Sam Altman will turn up at your doorstep the next day. As long as you aren’t sharing bank passwords or mentioning your real name, you can safely treat the large model as a close friend regarding current affairs commentary, psychological counselling, and the like. (Naturally, in this specific context, one wouldn’t touch domestic Chinese models with a bargepole due to censorship and privacy concerns). Every grumble receives a precise, earnest response—an affirmation akin to that of a soulmate. I fear no other place can offer such treatment.
When a tool stimulates a “confidant” so perfectly, we may abandon the search for real human connection, or cease striving to become individuals capable of independent thought and self-integration. Writing these words serves as a summary, but also as a reflection. The LLM itself is merely a technology; used well, it helps immensely, but used poorly, the consequences can be severe. The New York Times report “Chatbots Can Go Into a Delusional Spiral. Here’s How It Happens.” serves as a prime example. Although OpenAI has begun to address the issue of “sycophancy” in its models, I must acknowledge that while I can treat the model as a confidant and receive so-called “understanding and affirmation”, I cannot live permanently within that sensation. Joint studies from OpenAI and MIT found a positive correlation between ChatGPT usage and user loneliness: the more users utilised ChatGPT, the lonelier they felt.2 I could also argue that it is precisely because one is lonely that one turns to these LLMs, creating a vicious cycle. An LLM can serve as a seasoning for life, but it must never become the sole sustenance for the soul.
(To Be Continued)